These opt-in probes exercise real native discovery without creating a model thread or copying credentials. They are separate from the default unit suite.

node docker/context-profiles/native-probe.js
node docker/context-profiles/native-probe.js --claude
node docker/context-profiles/native-switch-probe.js
node docker/context-profiles/run-podman.js

The first command uses the locally installed Codex executable, a new private temporary home for each case, a local marketplace, and the native plugin cache. It starts a new app-server process and calls only initialize and skills/list. Lean, Lean with Angular's bundled resources, and Full excluding Python patterns must expose exactly their selected plugin skill names. Provider-owned system skills are reported separately. Every installed resource is checked against its source digest after removing the local marketplace's carrier source.

The Claude command uses the locally installed Claude executable, a private temporary home, empty setting sources, plugin validate, and plugin details with an inline plugin directory. It checks exact Lean/Full-with-exclusion skill inventories and zero agent, hook, MCP, and LSP components. Reported token costs are the provider's projections, not measured usage. Manifest attribution and version warnings remain visible.

The switch probe uses the product's managed store and isolated native adapter for Full, Lean, and rollback to Full. Preparation creates a separate provider home and registers the selected carrier, then opens a fresh app-server to verify discovery. Rollback first restores managed authority, then re-verifies the prior native home and selects it. The Full Python exclusion and unrelated bytes in the prior home must survive every transition. Each native pointer binds its managed store revision, carrier digest, exact provider version, and native executable SHA-256. Read-only status rechecks receipts, native configuration, cached resource bytes, and the pinned executable. Existing sessions and host registration remain unchanged.

The Podman runner runs the normal npm pack lifecycle, reports its archive SHA-256, and builds an isolated consumer from that archive. It installs runtime dependencies and pinned Codex 0.154.0 during the image build. The final container runs as the image's unprivileged node user, with networking disabled, all Linux capabilities dropped, no added host mounts, and no copied credentials. It checks all ten target/profile combinations through the packed public CLI and independent structural oracle, including exact carrier equality with the source checkout. It also checks the packed CLI's Full/Lean/rollback lifecycle, idempotency, stale revision rejection, Auto context loading, Suggest/Manual/dry-run boundaries, pinned receipt reuse, and no-workflow reset. It then repeats native Codex discovery and product native preparation/rollback. The packed CLI also prepares a native generation and verifies an isolated launch dry-run with no provider on PATH. Test helpers are copied separately into the image; they are not part of the published package.

An existing compatible Node image can be selected with ECC_CONTEXT_NODE_IMAGE=<image-id>. The default is node:22-bookworm-slim. The task image and private temporary build directory are removed afterward. Dependency download layers can remain in Podman's ordinary build cache. The runner never changes host harness configuration or mounts a host home.

The outcome evaluator (ai-eval.js) measures graded task success and provider usage across install arms; see ai-corpus.json for the 30-task repair corpus and complex-eval/DESIGN.md for the preregistered three-task complex-task benchmark (feature build, incident triage, security hardening) with scored hidden graders, reference solutions, and reproduction instructions.

These checks certify the observed discovery paths for the reported exact provider versions. They do not certify model invocation, skill workflow outcomes, implicit provider invocation of Auto, host activation, crash recovery, permission consent, or actual token savings. CLI-provided system skills still contribute to whole-session context.